CN110391022A - A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration - Google Patents
A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration Download PDFInfo
- Publication number
- CN110391022A CN110391022A CN201910673864.2A CN201910673864A CN110391022A CN 110391022 A CN110391022 A CN 110391022A CN 201910673864 A CN201910673864 A CN 201910673864A CN 110391022 A CN110391022 A CN 110391022A
- Authority
- CN
- China
- Prior art keywords
- resnet
- algorithm
- training
- mal
- ben
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 206010006187 Breast cancer Diseases 0.000 title claims abstract description 33
- 208000026310 Breast neoplasm Diseases 0.000 title claims abstract description 33
- 238000013135 deep learning Methods 0.000 title claims abstract description 31
- 230000001575 pathological effect Effects 0.000 title claims abstract description 28
- 238000013508 migration Methods 0.000 title claims abstract description 22
- 230000005012 migration Effects 0.000 title claims abstract description 22
- 238000002405 diagnostic procedure Methods 0.000 title claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 60
- 238000000034 method Methods 0.000 claims abstract description 41
- 238000013526 transfer learning Methods 0.000 claims abstract description 31
- 238000003745 diagnosis Methods 0.000 claims abstract description 18
- 230000008569 process Effects 0.000 claims abstract description 14
- 238000012545 processing Methods 0.000 claims abstract description 14
- 238000012360 testing method Methods 0.000 claims abstract description 12
- 230000000505 pernicious effect Effects 0.000 claims abstract description 11
- 206010028980 Neoplasm Diseases 0.000 claims description 33
- 201000011510 cancer Diseases 0.000 claims description 15
- 201000010983 breast ductal carcinoma Diseases 0.000 claims description 11
- 210000002569 neuron Anatomy 0.000 claims description 10
- 206010048832 Colon adenoma Diseases 0.000 claims description 9
- 208000000265 Lobular Carcinoma Diseases 0.000 claims description 9
- 208000008771 Lymphadenopathy Diseases 0.000 claims description 9
- 241001440127 Phyllodes Species 0.000 claims description 9
- 208000013228 adenopathy Diseases 0.000 claims description 9
- 201000003714 breast lobular carcinoma Diseases 0.000 claims description 9
- 206010073096 invasive lobular breast carcinoma Diseases 0.000 claims description 9
- 210000001595 mastoid Anatomy 0.000 claims description 9
- 208000022271 tubular adenoma Diseases 0.000 claims description 9
- 208000009956 adenocarcinoma Diseases 0.000 claims description 8
- 201000010879 mucinous adenocarcinoma Diseases 0.000 claims description 8
- 230000006978 adaptation Effects 0.000 claims description 5
- 238000005192 partition Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 230000008676 import Effects 0.000 claims description 3
- 230000006872 improvement Effects 0.000 claims description 3
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 abstract description 21
- 201000010099 disease Diseases 0.000 abstract description 19
- 210000005075 mammary gland Anatomy 0.000 abstract description 6
- 230000007170 pathology Effects 0.000 abstract description 6
- 238000005070 sampling Methods 0.000 abstract description 4
- 238000001027 hydrothermal synthesis Methods 0.000 abstract 1
- 238000002474 experimental method Methods 0.000 description 8
- 238000000605 extraction Methods 0.000 description 7
- 230000008859 change Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 230000036541 health Effects 0.000 description 3
- 230000031016 anaphase Effects 0.000 description 2
- 238000004195 computer-aided diagnosis Methods 0.000 description 2
- 239000006071 cream Substances 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 210000004907 gland Anatomy 0.000 description 2
- 230000003211 malignant effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 208000030090 Acute Disease Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000003292 glue Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000004630 mental health Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000010827 pathological analysis Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30068—Mammography; Breast
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Quality & Reliability (AREA)
- Epidemiology (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Pathology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to technical field of image processing, disclose a kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration.The present invention realizes breast cancer pathological image subdivision auxiliary diagnosis using deep learning algorithm；ResNet-V2-152 is improved first, then constructs two modified hydrothermal process ResNet-Ben and ResNet-Mal, is respectively intended to four kinds of benign subclass diseases and four kinds of pernicious subclass diseases in identification mammary gland disease.To meet data demand, up-sampling equalization and data extending have been carried out to data before training.Multistage transfer learning mode is combined in the training process, and frozen crust is introduced to algorithm to reduce over-fitting risk when migrating in the first stage；During second segment transfer learning, knowledge sharing is carried out according to the soft migration thought of multi-task learning parameter.Final ResNet-Ben algorithm and ResNet-Mal algorithm achieve 96% or so test recognition accuracy.The present invention solves breast cancer pathology problem of image recognition using deep learning method.
Description
Technical field
The present invention relates to technical field of image processing, specifically, more particularly to a kind of depth based on multistage migration
Learn breast cancer pathological image and segments diagnostic method.
Background technique
Breast cancer is one of the main reason for causing women disease to be died.Estimate according to American Cancer Society, breast cancer is new within 2018
Morbidity number of cases 2,100,000 accounts for the 11.6% of the total cancer neopathy number of cases in the whole world in 2018, and there are about 630,000 people to die of cream for the whole world in 2018
Gland cancer.The disease incidence of global women with breast cancer increases year by year.For women, in 154 countries including China, mammary gland
Cancer disease incidence ranks first.Disease is promptly and accurately diagnosed and is played a crucial role to anaphase, although in recent years
Medical imaging technology development it is very considerable, but pathological image diagnosis still finally make a definite diagnosis, be classified and by stages in terms of play
Huge effect.The research of breast cancer pathological image subdivision diagnosis has more importantly meaning for anaphase.
For the diagosis of pathological image, mainly completed by veteran pathology expert, even and experience very
There is also serious mistaken diagnosis phenomenons by virologist abundant.Meanwhile pathology department's development of basic hospital is more slow, by various
The restriction of factor.The problems such as lacking regular training, equipment backwardness, opportunity to study less, leads to the leakage of basic hospital pathological diagnosis
Examine, misdiagnosis rate it is higher, this is all either unfavorable to the life and health of patient or mental health.Breast cancer pathological image
Computer-aided diagnosis is studied especially the medical and health care system in China to the backwoodsman level of medical and health of raising
Vital meaning.For mammary gland disease, mammary gland disease be divided into it is good pernicious, and under benign and malignant mammary gland disease
It is respectively present four kinds of subclass diseases again, is respectively as follows: adenopathy, fibrous tumours, Phyllode tumour, tubular adenoma, duct carcinoma, lobular carcinoma, glue
Sample cancer and mastoid process cancer.Diagnosis be in order to which doctor preferably treats, so be diagnosed to be mammary gland disease it is good it is pernicious after, as can
It is then more meaningful that diagnosis is finely divided to benign and malignant diseases.
Currently, to breast cancer pathology visual aids diagnosis research mainly have manual extraction feature traditional images processing with
Two methods of image procossing based on deep learning.Traditional images processing method needs artificial extraction characteristics of image, exists artificial
Extract the problems such as characteristic procedure complexity is high, test recognition accuracy is low and generalization ability is poor.Due to deep learning have it is extremely strong
The automatic Extracting Ability of feature, can solve conventional machines and learn dependence to manual extraction property.The present invention utilizes depth
Learning method solves breast cancer pathology problem of image recognition.
Summary of the invention
The low, medical image data according to the high final recognition accuracy of traditional images processing feature set forth above extraction complexity
It is unevenly distributed the problems such as insufficient with data volume, the present invention proposes one kind and be based on using breast cancer pathological image as class object
The deep learning breast cancer pathological image of multistage migration segments diagnostic method, the computer aided manufacturing of Lai Shixian breast cancer pathological image
Help diagnosis.
The technological means that the present invention uses is as follows:
A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration, includes the following steps:
S1, open source data set BreakHis is obtained, and it is pre-processed；
S2, building ResNet-v2-152 algorithm, and the transfer learning of first stage is carried out to ResNet-v2-152 algorithm；
S3, the ResNet-v2-152 algorithm for passing through first stage transfer learning is improved, obtains two improved algorithms
ResNet-Ben and ResNet-Mal；
S4, frozen crust is introduced to ResNet-Ben algorithm and ResNet-Mal algorithm；
S5, second is carried out to the ResNet-Ben and ResNet-Mal that introduce frozen crust using open source data set BreakHis
The transfer learning in stage；
S6, using open source data set BreakHis to by second stage transfer learning ResNet-Ben and ResNet-
Mal is finely adjusted training, completes the building to algorithm ResNet-Ben and ResNet-Mal.
Further, the preprocessing process in the step S1 specifically:
S11, using DC data as baseline, to the open source data set BreakHis carry out equalization processing；
S12, by the modes such as random cropping and rotation to the open source data set BreakHis after equalization processing again
Carry out expansion processing；
S13, will expand treated open source data set BreakHis using the ratio cut partition of 9:1 as training set and test set
It closes；
S14, the format that training set is converted into TFRecord are convenient to carry out batch reading to data in the training process.
Further, in the step S2 first stage transfer learning specifically:
S21, ResNet-v2-152 algorithm is built using tensorflow frame；
S22, pre-training is carried out to ResNet-v2-152 algorithm using ImgeNet data set, and exports pre-training
The pre-training parameter of ResNet-v2-152 conventional part.
Further, the ResNet-v2-152 algorithm of first stage transfer learning, tool are passed through in the improvement in the step S3
Body are as follows:
S31, remove the softmax layer being made of in ResNet-v2-152 1000 neurons, and in first full connection
The adaptation layer being made of 512 neurons and the classification layer being made of 4 neurons are added after layer；
S32, by using the method in above-mentioned steps S31, construct two improved ResNet-v2-152 algorithms, respectively
For ResNet-Ben algorithm and ResNet-Mal algorithm；ResNet-Ben algorithm be used to identify four kinds of benign subclass data,
ResNet-Mal algorithm is used to identify four kinds of pernicious subclass data；
S33, pre-training parameter derived in step S22 is imported into ResNet-Ben algorithm and ResNet-Mal algorithm.
Further, described four kinds benign subclass data include adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma；It is described
Four kinds of pernicious subclass data include duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer.
Further, frozen crust is introduced to ResNet-Ben and ResNet-Mal in the step S4, specifically:
S41, will import pre-training parameter ResNet-Ben algorithm and ResNet-Mal algorithm first convolution module
Parameter is carried out with first residual error module to freeze.
Further, in the step S5 second stage transfer learning, specifically:
S51, the ResNet-Ben for introducing frozen crust is calculated using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set
Method is trained；
S52, using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set to introduce frozen crust ResNet-Mal
Algorithm is trained.
Further, the fine tuning training in the step S6, specifically:
S61, ResNet-Ben algorithm is finely adjusted using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set
Training；
S62, instruction is finely adjusted to ResNet-Mal algorithm using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set
Practice.
Compared with the prior art, the invention has the following advantages that
1, the deep learning breast cancer pathological image provided by the invention based on multistage migration segments diagnostic method, utilizes
Improved deep learning algorithm realizes the subdivision of eight kinds of subclass acute diseases of breast cancer pathological image, solves traditional medical image recognition
Method characteristic extraction process complexity height and the low problem of final recognition accuracy.
2, the deep learning breast cancer pathological image provided by the invention based on multistage migration segments diagnostic method, uses
Data extending and multistage transfer learning method solve the problems, such as that data volume is insufficient in deep learning medical image identification process,
Feasible scheme is provided for computer-aided diagnosis.Solves data set not using top sampling method in data set pretreatment
Equalization problem is realized data set in the way of random cropping and rotation etc. and expanded.
3, the deep learning breast cancer pathological image provided by the invention based on multistage migration segments diagnostic method, is instructing
Multistage transfer learning is introduced during practicing, deep learning algorithm is identified in natural image and led by first stage transfer learning
The knowledge and mode learnt on domain is applied to medical field, frozen crust is introduced in first stage migration, due to deep learning
With very strong geometric invariance, the introducing of frozen crust can reduce training difficulty under the premise of ensureing accuracy, reduce
The risk of fitting is inspired in second stage transfer learning by multi-task learning thought, promotes itself using the data set of other side
Recognition accuracy, realize knowledge sharing, improve the Generalization Capability of algorithm.
4, the deep learning breast cancer pathological image provided by the invention based on multistage migration segments diagnostic method, right
ResNet-v2-152 algorithm improves, and introduces adaptation layer and four classification layers, and the introducing of adaptation layer can be realized preferably finally
Small parameter perturbations, the introducing for layer of classifying can complete point of benign four kinds of subclass diseases and pernicious four kinds of subclass diseases well
Class.
To sum up, applying the technical scheme of the present invention, it is quasi- to solve the high final identification of traditional images processing feature extraction complexity
Exactness is low, medical image data is unevenly distributed the problems such as insufficient with data volume.
The present invention can be widely popularized in fields such as image procossings based on the above reasons.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to do simply to introduce, it should be apparent that, the accompanying drawings in the following description is this hair
Bright some embodiments for those of ordinary skill in the art without any creative labor, can be with
It obtains other drawings based on these drawings.
Fig. 1 is the method for the present invention flow chart.
Fig. 2 is that ResNet-Ben provided in an embodiment of the present invention finally finely tunes training accuracy change curve.
Fig. 3 is that ResNet-Ben provided in an embodiment of the present invention finally finely tunes training loss change curve.
Fig. 4 ResNet-Mal provided in an embodiment of the present invention finally finely tunes training accuracy change curve.
Fig. 5 ResNet-Mal provided in an embodiment of the present invention finally finely tunes training loss change curve.
Specific embodiment
In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction in the embodiment of the present invention
Attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
The embodiment of a part of the invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
The model that the present invention protects all should belong in member's every other embodiment obtained without making creative work
It encloses.
It should be noted that description and claims of this specification and term " first " in above-mentioned attached drawing, "
Two " etc. be to be used to distinguish similar objects, without being used to describe a particular order or precedence order.It should be understood that using in this way
Data be interchangeable under appropriate circumstances, so as to the embodiment of the present invention described herein can in addition to illustrating herein or
Sequence other than those of description is implemented.In addition, term " includes " and " having " and their any deformation, it is intended that cover
Cover it is non-exclusive include, for example, the process, method, system, product or equipment for containing a series of steps or units are not necessarily limited to
Step or unit those of is clearly listed, but may include be not clearly listed or for these process, methods, product
Or other step or units that equipment is intrinsic.
If Fig. 1 shows, the present invention provides a kind of deep learning breast cancer pathological image subdivisions based on multistage migration to examine
Disconnected method, includes the following steps:
S1, open source data set BreakHis is obtained, and it is pre-processed；
Preprocessing process in step S1 specifically:
There are the unbalanced problem of serious data volume, duct carcinoma data (DC) are obvious on the high side easy for S11, the data got
It causes network to duct carcinoma data overlearning, is brought to solve the problems, such as that data are unbalanced, in the present embodiment, be with DC data
Baseline, split set of source data BreakHis carries out equalization processing in a manner of up-sampling；
S12, artificial intelligence are the intelligence under big data driving, to solve shortage of data bring overfitting problem, this reality
It applies in example, the open source data set BreakHis after equalization processing is expanded again by the modes such as random cropping and rotation
Fill processing；
S13, will expand treated open source data set BreakHis using the ratio cut partition of 9:1 as training set and test set
It closes；
S14, the format that training set is converted into TFRecord are convenient to carry out batch reading to data in the training process.
S2, building ResNet-v2-152 algorithm, and the transfer learning of first stage is carried out to ResNet-v2-152 algorithm；
Transfer learning migration can have monitoring data or the structure of knowledge, improve target domain from related or near field
The learning effect of task.The transfer learning of first stage in step S2 specifically:
S21, ResNet-v2-152 algorithm is built using tensorflow frame；
S22, pre-training is carried out to ResNet-v2-152 algorithm using ImgeNet data set, and exports pre-training
The pre-training parameter of ResNet-v2-152 conventional part.
S3, the ResNet-v2-152 algorithm for passing through first stage transfer learning is improved, obtains two improved algorithms
ResNet-Ben and ResNet-Mal；
For the subdivision auxiliary diagnosis for completing this eight kinds of subclass disease (benign four kinds, pernicious four kinds, totally eight kinds), directly training
The deep learning algorithm of eight classification, but the more calculation amounts of classification classified are bigger, also will affect the accuracy of classification
And generalization.Eight classification tasks are then divided into the subtask of two four classification, one of them task is completed good in the present embodiment
The subdivision diagnosis of temper class disease, in addition a task completes pernicious subclass disease subdivision diagnosis.Original ResNet-v2-152 is
In order to solve thousand classification problem of ImageNet, the full articulamentum of the last layer is the softmax for having 1000 neurons
Layer, to better solve breast cancer disease Neo-Confucianism image classification problem, in the present embodiment, the first stage is passed through in the improvement in step S3
The ResNet-v2-152 algorithm of transfer learning, specifically:
S31, remove the softmax layer being made of in ResNet-v2-152 1000 neurons, and in first full connection
The adaptation layer being made of 512 neurons and the classification layer being made of 4 neurons are added after layer；
S32, by using the method in above-mentioned steps S31, construct two improved ResNet-v2-152 algorithms, respectively
For ResNet-Ben algorithm and ResNet-Mal algorithm；ResNet-Ben algorithm be used to identify four kinds of benign subclass data,
ResNet-Mal algorithm is used to identify four kinds of pernicious subclass data；
S33, pre-training parameter derived in step S22 is imported into ResNet-Ben algorithm and ResNet-Mal algorithm.
Since deep learning has very strong geometry indeformable, shallow-layer convolution extracts shallow-layer characteristic, and deep layer network extracts more
Abstract semantic information, and for deep learning problem of image recognition, shallow-layer network is to natural image and medical image
Feature extraction mode is all identical.Then in the present embodiment,
S4, frozen crust is introduced to ResNet-Ben algorithm and ResNet-Mal algorithm；Pre-training parameter will be imported
First convolution module and first residual error module of ResNet-Ben algorithm and ResNet-Mal algorithm carry out parameter and freeze.Freeze
The introducing for tying layer can reduce trained difficulty while guaranteeing accuracy, reduce over-fitting risk.
The present invention will complete the classification of benign four kinds of subclass diseases and pernicious four kinds of subclass diseases, although and this two tasks
Difference, but have similarity.It is inspired by multi-task learning thought, the present invention proposes a kind of cream based on multistage transfer learning
Gland cancer pathological image classification method is finally carrying out parameter to the algorithm for solving itself task with the training data in itself task
Before fine tuning, the algorithm for solving itself task is trained first with the training data of other side, to realize knowledge sharing, promotes final calculate
The Generalization Capability of method model.I.e. in the present embodiment, specifically:
S5, second is carried out to the ResNet-Ben and ResNet-Mal that introduce frozen crust using open source data set BreakHis
The transfer learning in stage；
The transfer learning of second stage in step S5, specifically:
S51, the ResNet-Ben for introducing frozen crust is calculated using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set
Method is trained；
S52, using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set to introduce frozen crust ResNet-Mal
Algorithm is trained.(transfer training can also be carried out using other breast cancer pathology image data sets in this part such as to open
Set of source data TMA)
S6, using open source data set BreakHis to by second stage transfer learning ResNet-Ben and ResNet-
Mal is finely adjusted training, completes the building to algorithm ResNet-Ben and ResNet-Mal.
Fine tuning training in step S6, specifically:
S61, ResNet-Ben algorithm is finely adjusted using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set
Training；
S62, instruction is finely adjusted to ResNet-Mal algorithm using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set
Practice.
Fig. 1 to Fig. 4 is finally to finely tune the loss of quasi- training process and the change curve of accuracy, in the training process present invention
20 are set by batch-size, initial learning rate is 0.01, and is decayed in the training process to learning rate, final micro-
It adjusts in training, ResNet-Mal and ResNet-Ben is had trained 3000 times, it can be seen from the figure that at iteration 1500 times or so
When training accuracy just already close to 1, illustrating method mentioned in the present invention well reduces trained difficulty, also drops
Low dependence of the algorithm to target data, and possess very high trained accuracy.
Embodiment
In the present embodiment, using Google tensorflow as deep learning frame, in 64 Ubuntu18.04 operation systems
Deep learning CNN network model is built under system environment.Hardware uses Intel i7-7800X processor and GeForce RTX
2080Ti video card accelerates training.The ratio cut partition of logarithm 9:1 accordingly has gone out training set and test set after data prediction.In reality
The each stage tested has carried out accuracy test, calculates the numerical value got as Accuracy evaluation index using formula (1).
Wherein, BenrightIndicate that ResNet-Ben identifies correct picture number, MalrightIndicate ResNet-Mal identification
Correct picture number, NtotalIndicate that the sum of the test set marked off, ACC are the index for evaluating inventive energy.
Performance of the invention is analyzed below by following three experiments
Embodiment 1
Experiment one: using initial data (data set for not equalized and being expanded), to importing ImageNet pre-training
The ResNet-Ben and ResNet-Mal of ckpt is trained, and tests training result.
Embodiment 2
Experiment two: (carrying out up-sampling equalization to initial data and data set expand) using pretreated data set,
The ResNet-Ben and ResNet-Mal that import ImageNet pre-training ckpt are trained, and test training result.
Embodiment 3
Experiment three: pre-processing data, two stages transfer learning is combined in training process, and test training result.
Experimental results show is in table 1
One experimental result of table
| Experiment | Test accuracy |
| Experiment one | 86.6±0.2 |
| Experiment two | 95.5±0.2 |
| Experiment three | 96.0±0.2 |
As can be seen from Table I, data balancing and data extending are made that tremendous contribution to the promotion of accuracy, simultaneously
As the expansion accuracy of transfer learning also has apparent promotion, the feasibility of the method for the present invention is sufficiently demonstrated.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
In the above embodiment of the invention, it all emphasizes particularly on different fields to the description of each embodiment, does not have in some embodiment
The part of detailed description, reference can be made to the related descriptions of other embodiments.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement；And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme.
Claims (8)
1. a kind of deep learning breast cancer pathological image based on multistage migration segments diagnostic method, which is characterized in that including
Following steps:
S1, open source data set BreakHis is obtained, and it is pre-processed；
S2, building ResNet-v2-152 algorithm, and the transfer learning of first stage is carried out to ResNet-v2-152 algorithm；
S3, the ResNet-v2-152 algorithm for passing through first stage transfer learning is improved, obtains two improved algorithms
ResNet-Ben and ResNet-Mal；
S4, frozen crust is introduced to ResNet-Ben algorithm and ResNet-Mal algorithm；
S5, second stage is carried out to the ResNet-Ben and ResNet-Mal that introduce frozen crust using open source data set BreakHis
Transfer learning；
S6, using open source data set BreakHis to by second stage transfer learning ResNet-Ben and ResNet-Mal into
Row fine tuning training, completes the building to algorithm ResNet-Ben and ResNet-Mal.
2. the deep learning breast cancer pathological image according to claim 1 based on multistage migration segments diagnostic method,
It is characterized in that, the preprocessing process in the step S1 specifically:
S11, using DC data as baseline, to the open source data set BreakHis carry out equalization processing；
S12, the open source data set BreakHis after equalization processing is carried out again by modes such as random cropping and rotations
Expansion processing；
S13, will expand treated open source data set BreakHis using the ratio cut partition of 9:1 be training set and test set；
S14, the format that training set is converted into TFRecord are convenient to carry out batch reading to data in the training process.
3. the deep learning breast cancer pathological image according to claim 1 or 2 based on multistage migration segments diagnosis side
Method, which is characterized in that the transfer learning of first stage in the step S2 specifically:
S21, ResNet-v2-152 algorithm is built using tensorflow frame；
S22, pre-training is carried out to ResNet-v2-152 algorithm using ImgeNet data set, and exports the ResNet- of pre-training
The pre-training parameter of v2-152 conventional part.
4. the deep learning breast cancer pathological image according to claim 1 or 2 based on multistage migration segments diagnosis side
Method, which is characterized in that the ResNet-v2-152 algorithm of first stage transfer learning is passed through in the improvement in the step S3, specifically
Are as follows:
S31, remove the softmax layer being made of in ResNet-v2-152 1000 neurons, and after first full articulamentum
The adaptation layer being made of 512 neurons and the classification layer being made of 4 neurons is added；
S32, by using the method in above-mentioned steps S31, construct two improved ResNet-v2-152 algorithms, respectively
ResNet-Ben algorithm and ResNet-Mal algorithm；ResNet-Ben algorithm is used to identify four kinds of benign subclass data, ResNet-
Mal algorithm is used to identify four kinds of pernicious subclass data；
S33, pre-training parameter derived in step S22 is imported into ResNet-Ben algorithm and ResNet-Mal algorithm.
5. the deep learning breast cancer pathological image according to claim 4 based on multistage migration segments diagnostic method,
It is characterized in that, four kinds of benign subclass data include adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma；Four kinds of evils
Temper class data include duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer.
6. the deep learning breast cancer pathological image according to claim 1 or 2 based on multistage migration segments diagnosis side
Method, which is characterized in that frozen crust is introduced to ResNet-Ben and ResNet-Mal in the step S4, specifically:
S41, will import pre-training parameter ResNet-Ben algorithm and ResNet-Mal algorithm first convolution module and the
One residual error module carries out parameter and freezes.
7. the deep learning breast cancer pathological image according to claim 1 or 2 based on multistage migration segments diagnosis side
Method, which is characterized in that the transfer learning of second stage in the step S5, specifically:
S51, using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set to introduce frozen crust ResNet-Ben algorithm into
Row training；
S52, using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set to introduce frozen crust ResNet-Mal algorithm
It is trained.
8. the deep learning breast cancer pathological image according to claim 1 or 2 based on multistage migration segments diagnosis side
Method, which is characterized in that the fine tuning training in the step S6, specifically:
S61, training is finely adjusted to ResNet-Ben algorithm using adenopathy, fibrous tumours, Phyllode tumour and tubular adenoma data set；
S62, training is finely adjusted to ResNet-Mal algorithm using duct carcinoma, lobular carcinoma, mucinous carcinoma and mastoid process cancer data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910673864.2A CN110391022A (en) | 2019-07-25 | 2019-07-25 | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910673864.2A CN110391022A (en) | 2019-07-25 | 2019-07-25 | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110391022A true CN110391022A (en) | 2019-10-29 |
Family
ID=68287157
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910673864.2A Pending CN110391022A (en) | 2019-07-25 | 2019-07-25 | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110391022A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110889838A (en) * | 2019-11-26 | 2020-03-17 | 武汉纺织大学 | Fabric defect detection method and device |
| CN112086197A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学附属翔安医院 | Mammary nodule detection method and system based on ultrasonic medicine |
| WO2021143781A1 (en) * | 2020-01-14 | 2021-07-22 | 之江实验室 | Multi-center synergetic cancer prognosis prediction system based on multi-source migration learning |
| CN113486917A (en) * | 2021-05-17 | 2021-10-08 | 西安电子科技大学 | Radar HRRP small sample target identification method based on metric learning |
| CN114820568A (en) * | 2022-05-20 | 2022-07-29 | 青岛农业大学 | Method and equipment for building corn leaf disease identification model and storage medium |
-
2019
- 2019-07-25 CN CN201910673864.2A patent/CN110391022A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110889838A (en) * | 2019-11-26 | 2020-03-17 | 武汉纺织大学 | Fabric defect detection method and device |
| WO2021143781A1 (en) * | 2020-01-14 | 2021-07-22 | 之江实验室 | Multi-center synergetic cancer prognosis prediction system based on multi-source migration learning |
| US11456078B2 (en) | 2020-01-14 | 2022-09-27 | Zhejiang Lab | Multi-center synergetic cancer prognosis prediction system based on multi-source migration learning |
| CN112086197A (en) * | 2020-09-04 | 2020-12-15 | 厦门大学附属翔安医院 | Mammary nodule detection method and system based on ultrasonic medicine |
| CN113486917A (en) * | 2021-05-17 | 2021-10-08 | 西安电子科技大学 | Radar HRRP small sample target identification method based on metric learning |
| CN113486917B (en) * | 2021-05-17 | 2023-06-02 | 西安电子科技大学 | Radar HRRP small sample target recognition method based on metric learning |
| CN114820568A (en) * | 2022-05-20 | 2022-07-29 | 青岛农业大学 | Method and equipment for building corn leaf disease identification model and storage medium |
| CN114820568B (en) * | 2022-05-20 | 2024-04-30 | 青岛农业大学 | Corn leaf disease identification model building method, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110391022A (en) | A kind of deep learning breast cancer pathological image subdivision diagnostic method based on multistage migration | |
| CN108062753B (en) | Unsupervised domain self-adaptive brain tumor semantic segmentation method based on deep counterstudy | |
| CN108596882B (en) | The recognition methods of pathological picture and device | |
| CN110889853B (en) | Tumor segmentation method based on residual error-attention deep neural network | |
| Yang et al. | Class-balanced deep neural network for automatic ventricular structure segmentation | |
| CN109389584A (en) | Multiple dimensioned rhinopharyngeal neoplasm dividing method based on CNN | |
| CN109063710A (en) | Based on the pyramidal 3D CNN nasopharyngeal carcinoma dividing method of Analysis On Multi-scale Features | |
| CN109191476A (en) | The automatic segmentation of Biomedical Image based on U-net network structure | |
| CN110675411B (en) | Cervical squamous intraepithelial lesion recognition algorithm based on deep learning | |
| CN108648182B (en) | Breast cancer nuclear magnetic resonance image tumor region segmentation method based on molecular subtype | |
| CN113808146B (en) | Multi-organ segmentation method and system for medical image | |
| WO2023045231A1 (en) | Method and apparatus for facial nerve segmentation by decoupling and divide-and-conquer | |
| CN101609485B (en) | Medical image diagnosing system and diagnosing method based on migrated nuclear matching tracing | |
| CN116309368B (en) | Lung cancer pathological diagnosis system based on deep migration learning | |
| Nazki et al. | Image-to-image translation with GAN for synthetic data augmentation in plant disease datasets | |
| CN109215040A (en) | A kind of tumor of breast dividing method based on multiple dimensioned weighting study | |
| Yonekura et al. | Improving the generalization of disease stage classification with deep CNN for glioma histopathological images | |
| CN112381846A (en) | Ultrasonic thyroid nodule segmentation method based on asymmetric network | |
| CN113555087A (en) | Artificial intelligence film reading method based on convolutional neural network algorithm | |
| Chakraborty et al. | Detection of cancerous tissue in histopathological images using Dual-Channel Residual Convolutional Neural Networks (DCRCNN) | |
| CN115471701A (en) | Lung adenocarcinoma histology subtype classification method based on deep learning and transfer learning | |
| Bansal et al. | An improved hybrid classification of brain tumor MRI images based on conglomeration feature extraction techniques | |
| Song et al. | Red blood cell classification based on attention residual feature pyramid network | |
| CN112330662B (en) | Medical image segmentation system and method based on multi-level neural network | |
| CN112419335B (en) | Shape loss calculation method of cell nucleus segmentation network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191029 |
|
| RJ01 | Rejection of invention patent application after publication |